4 Methods and Material
4.8 Data analyses
4.8.1 Statistical procedures for Study I
In study I, a representative sample consisting of 11,956 adolescents (M/F: 5225/6731) with a mean age of 14.9 ± 0.89 years were included in the study. In the respective analyses, both categorical and dimensional evaluations using the YDQ were studied independently. Descriptive analysis of the prevalence for AIU, MIU and PIU stratified by gender and country were conducted. Statistical significance between group proportions were analysed using the Bonferroni-adjusted Wald test after a multinomial regression model, with Internet user group as the dependent variable and country as the independent variable. Mean hours online per-day were calculated and stratified by gender and Internet user group using a two-factor analysis of variance (ANOVA). Goodman and Kruskal’s Gamma test was utilized to explore the association between Internet user groups and online activities distinctly for males and females. Gender differences were compared using likelihood ratio tests. In the multinomial regression analysis examining the association between PIU and social factors, Internet user groups were entered as the dependent variable, while social factors were entered as the independent variables. In order to corroborate results from the categorical analysis using the multinomial regression model, a linear regression model was employed using the YDQ total score as a continuous variable. In the 16 variables used in the regression analysis, at least one or more missing values were detected. This yielded a total of 21.2 percent of the sample (n=2534) that had at least one or more missing values in one of the respective sixteen variables. In order to prevent estimation bias by excluding these subjects, missing values were imputed using the multivariate imputation by chained equations (MICE) algorithm. Regression models were then calculated for the imputed data sets and results were combined. Relative Risk Ratios (RRR) and regression coefficients (β) with 95% confidence intervals (CI) and p-values were reported. A critical value of p < 0.05 was considered to be statistically significant for all models.
4.8.2 Statistical procedures for Study II
Study II is a systematic review and meta-analysis with the aim to methodically evaluate previous research in order to ascertain the relationship between PIU and comorbid psychopathology. An electronic literature search of key indicators was conducted using the following databases: MEDLINE, PsycARTICLES, PsychINFO, Global Health, and Web of Science. There were no restrictions on language, time or publication status. Articles were systematically and independently reviewed by the authors. An impartial evaluation regarding the study type, study population, methodology, outcome measures, effect sizes and interpretation of results were conducted. The inclusion criteria for studies involved population-based studies with a large sample size (n ≥ 1200 subjects), legitimate diagnostic criteria for PIU, subsequent reporting on the correlation between PIU and predetermined psychopathologies, and the psychometric outcome measures assessing psychopathology. This yielded a total of 20 studies for the analysis. The respective twenty articles were rated according to the scheme proposed by the Oxford Centre for Evidence-Based Medicine Results and evaluated by using the following criteria: observation of a full or partial association, significance level, and adjustments for confounders. Full association was considered when a correlation was found for both sexes after multivariate analyses. If a correlation was identified for only one gender, it was classified as a partial association. The geographical distribution of studies was also mapped.
Effect size of the associations was identified by either the original publication or calculated using the data of the respective publications. The calculated effect sizes were either Cohen’s d or R2. In order to
compare the different associations, the effect sizes d and R2 were stated as small, moderate and large in
accordance to Cohen d, while ORs were converted into the respective groups in accordance to Chinn. The effect sizes were interpreted as small (d=0.2, R2=0.01, OR=1.45), moderate (d=0.5, R2=0.06,
OR=2.50) and large (d=0.8, R2=0.14, OR=4.25). The potential effect of publication bias was also
assessed by using a funnel plot graph and tested using Egger’s linear regression method. If publication bias was found, a trim and fill method was used to estimate the number of missing studies and adjusted accordingly.
psychopathology. In Study V, a longitudinal design was used to ascertain the preventive effect of mental health action in schools (MHAS) on reducing the onset risk of PIU and psychosocial impairments.
4.8 DATA ANALYSES
4.8.1 Statistical procedures for Study I
In study I, a representative sample consisting of 11,956 adolescents (M/F: 5225/6731) with a mean age of 14.9 ± 0.89 years were included in the study. In the respective analyses, both categorical and dimensional evaluations using the YDQ were studied independently. Descriptive analysis of the prevalence for AIU, MIU and PIU stratified by gender and country were conducted. Statistical significance between group proportions were analysed using the Bonferroni-adjusted Wald test after a multinomial regression model, with Internet user group as the dependent variable and country as the independent variable. Mean hours online per-day were calculated and stratified by gender and Internet user group using a two-factor analysis of variance (ANOVA). Goodman and Kruskal’s Gamma test was utilized to explore the association between Internet user groups and online activities distinctly for males and females. Gender differences were compared using likelihood ratio tests. In the multinomial regression analysis examining the association between PIU and social factors, Internet user groups were entered as the dependent variable, while social factors were entered as the independent variables. In order to corroborate results from the categorical analysis using the multinomial regression model, a linear regression model was employed using the YDQ total score as a continuous variable. In the 16 variables used in the regression analysis, at least one or more missing values were detected. This yielded a total of 21.2 percent of the sample (n=2534) that had at least one or more missing values in one of the respective sixteen variables. In order to prevent estimation bias by excluding these subjects, missing values were imputed using the multivariate imputation by chained equations (MICE) algorithm. Regression models were then calculated for the imputed data sets and results were combined. Relative Risk Ratios (RRR) and regression coefficients (β) with 95% confidence intervals (CI) and p-values were reported. A critical value of p < 0.05 was considered to be statistically significant for all models.
4.8.2 Statistical procedures for Study II
Study II is a systematic review and meta-analysis with the aim to methodically evaluate previous research in order to ascertain the relationship between PIU and comorbid psychopathology. An electronic literature search of key indicators was conducted using the following databases: MEDLINE, PsycARTICLES, PsychINFO, Global Health, and Web of Science. There were no restrictions on language, time or publication status. Articles were systematically and independently reviewed by the authors. An impartial evaluation regarding the study type, study population, methodology, outcome measures, effect sizes and interpretation of results were conducted. The inclusion criteria for studies involved population-based studies with a large sample size (n ≥ 1200 subjects), legitimate diagnostic criteria for PIU, subsequent reporting on the correlation between PIU and predetermined psychopathologies, and the psychometric outcome measures assessing psychopathology. This yielded a total of 20 studies for the analysis. The respective twenty articles were rated according to the scheme proposed by the Oxford Centre for Evidence-Based Medicine Results and evaluated by using the following criteria: observation of a full or partial association, significance level, and adjustments for confounders. Full association was considered when a correlation was found for both sexes after multivariate analyses. If a correlation was identified for only one gender, it was classified as a partial association. The geographical distribution of studies was also mapped.
Effect size of the associations was identified by either the original publication or calculated using the data of the respective publications. The calculated effect sizes were either Cohen’s d or R2. In order to
compare the different associations, the effect sizes d and R2 were stated as small, moderate and large in
accordance to Cohen d, while ORs were converted into the respective groups in accordance to Chinn. The effect sizes were interpreted as small (d=0.2, R2=0.01, OR=1.45), moderate (d=0.5, R2=0.06,
OR=2.50) and large (d=0.8, R2=0.14, OR=4.25). The potential effect of publication bias was also
assessed by using a funnel plot graph and tested using Egger’s linear regression method. If publication bias was found, a trim and fill method was used to estimate the number of missing studies and adjusted accordingly.
4.8.3 Statistical procedures for Study III
In study III, a representative sample of 11,356 school-based adolescents (M/F: 4856/6500) with a mean age of 14.9 years were included in the analyses. Symptoms of depression were assessed using the Beck Depression Inventory-II (BDI-II). Anxiety was assessed using the Zung Self-Rating Anxiety Scale (Z-SAS). Emotional symptoms, conduct problems, hyperactivity-inattention, peer relationship problems and pro-social behaviours were measured using the Strengths and Difficulties Questionnaire (SDQ). Self-injurious behaviours (SIB) were evaluated using the Deliberate Self-Harm Inventory (DSHI). Suicidal ideation and suicide attempts were assessed using the Paykel Suicide Scale (PSS).
The prevalence of psychopathology and self-destructive behaviours were calculated and stratified by Internet user group. To analyse the relationship between PIU, psychopathology and self-destructive behaviours, a multilevel mixed-effects linear regression analysis was performed. In this model, PIU was entered as the dependent variable, while age, gender, psychopathological scores (BDI-II, Z-SAS, SDQ) and categorical self-destructive behaviours (DSHI, PSS) were entered as level 1 fixed effects, school as level 2 random intercept and country as level 3 random intercept. The estimation method was full ML with independent covariance structure. A subsequent stepwise reduction of the regression model was conducted in order to minimize the Bayes Information Criterion (BIC). Potential interactions of the predictors with gender and country were also explored. Each predictor was analysed as a separate model. PIU was used as the dependent variable, while the predictor, gender, country and interaction terms (predictor*gender) and (predictor*country) were entered as level 1 fixed effects, with school as level 2 random intercept. To avoid estimation problems, we combined the categories ‘suicidal ideation’ and ‘suicide attempts’ into one category discerned as ‘suicidal behaviours’. Regression coefficients (β) with 95% confidence intervals (CI) and p-values were reported for the respective models. A critical value of p < 0.05 was considered to be statistically significant for all models.
4.8.4 Statistical procedures for Study IV
In study IV, a representative sample of 11,196 school-based adolescents comprising 43.1 percent males and 56.9 percent females (M/F: 4830/6366) with a mean age of 14.88 ± 0.88 years were included in the analysis. Data on health risk-behaviours were obtained by using questions procured from the Global School-Based Student Health Survey (GSHS). The prevalence of individual risk-behaviours among Internet user groups was calculated for males and females. To ascertain statistically significant differences between group proportions, multiple pairwise comparisons using the two-sided z-test with Bonferroni adjusted p-values was performed.
Extended analyses were conducted to test the effect of individual risk-behaviours on MIU and PIU using generalized linear mixed models (GLMM) with a multinomial logit link and full maximum likelihood estimation. In the GLMM analysis, MIU and PIU were entered as the outcome measures with AIU as the reference category, individual risk-behaviours were entered as Level 1 fixed effects, school as Level 2 random intercept and country as Level 3 random intercept. Variance components were used as the covariance structure for the random effects. To study the moderating effect of gender, interaction terms (gender*risk-behaviour) were fitted into the regression model. Adjustments for age and gender were applied to relevant GLMM models. Odds ratios (OR) with 95% confidence intervals (CI) and p- values were reported for the respective models.
In the analysis on multiple risk-behaviours, the mean (M) and standard error of the mean (SEM) were calculated for Internet user groups and stratified by gender. Box and whisker plots were used to illustrate these relationships. Statistical significance between multiple risk-behaviours and gender was assessed using independent samples t-test. One-way analyses of variance (ANOVA) with post hoc pairwise comparisons were employed to assess the statistical significance between multiple risk- behaviours and Internet user groups. Moreover, a regression variable plot was conducted to elucidate the linear relationship between the number of hours online per day and the number of risk-behaviours by Internet user group. A critical value of p < 0.05 was considered to be statistically significant for all models.
4.8.3 Statistical procedures for Study III
In study III, a representative sample of 11,356 school-based adolescents (M/F: 4856/6500) with a mean age of 14.9 years were included in the analyses. Symptoms of depression were assessed using the Beck Depression Inventory-II (BDI-II). Anxiety was assessed using the Zung Self-Rating Anxiety Scale (Z-SAS). Emotional symptoms, conduct problems, hyperactivity-inattention, peer relationship problems and pro-social behaviours were measured using the Strengths and Difficulties Questionnaire (SDQ). Self-injurious behaviours (SIB) were evaluated using the Deliberate Self-Harm Inventory (DSHI). Suicidal ideation and suicide attempts were assessed using the Paykel Suicide Scale (PSS).
The prevalence of psychopathology and self-destructive behaviours were calculated and stratified by Internet user group. To analyse the relationship between PIU, psychopathology and self-destructive behaviours, a multilevel mixed-effects linear regression analysis was performed. In this model, PIU was entered as the dependent variable, while age, gender, psychopathological scores (BDI-II, Z-SAS, SDQ) and categorical self-destructive behaviours (DSHI, PSS) were entered as level 1 fixed effects, school as level 2 random intercept and country as level 3 random intercept. The estimation method was full ML with independent covariance structure. A subsequent stepwise reduction of the regression model was conducted in order to minimize the Bayes Information Criterion (BIC). Potential interactions of the predictors with gender and country were also explored. Each predictor was analysed as a separate model. PIU was used as the dependent variable, while the predictor, gender, country and interaction terms (predictor*gender) and (predictor*country) were entered as level 1 fixed effects, with school as level 2 random intercept. To avoid estimation problems, we combined the categories ‘suicidal ideation’ and ‘suicide attempts’ into one category discerned as ‘suicidal behaviours’. Regression coefficients (β) with 95% confidence intervals (CI) and p-values were reported for the respective models. A critical value of p < 0.05 was considered to be statistically significant for all models.
4.8.4 Statistical procedures for Study IV
In study IV, a representative sample of 11,196 school-based adolescents comprising 43.1 percent males and 56.9 percent females (M/F: 4830/6366) with a mean age of 14.88 ± 0.88 years were included in the analysis. Data on health risk-behaviours were obtained by using questions procured from the Global School-Based Student Health Survey (GSHS). The prevalence of individual risk-behaviours among Internet user groups was calculated for males and females. To ascertain statistically significant differences between group proportions, multiple pairwise comparisons using the two-sided z-test with Bonferroni adjusted p-values was performed.
Extended analyses were conducted to test the effect of individual risk-behaviours on MIU and PIU using generalized linear mixed models (GLMM) with a multinomial logit link and full maximum likelihood estimation. In the GLMM analysis, MIU and PIU were entered as the outcome measures with AIU as the reference category, individual risk-behaviours were entered as Level 1 fixed effects, school as Level 2 random intercept and country as Level 3 random intercept. Variance components were used as the covariance structure for the random effects. To study the moderating effect of gender, interaction terms (gender*risk-behaviour) were fitted into the regression model. Adjustments for age and gender were applied to relevant GLMM models. Odds ratios (OR) with 95% confidence intervals (CI) and p- values were reported for the respective models.
In the analysis on multiple risk-behaviours, the mean (M) and standard error of the mean (SEM) were calculated for Internet user groups and stratified by gender. Box and whisker plots were used to illustrate these relationships. Statistical significance between multiple risk-behaviours and gender was assessed using independent samples t-test. One-way analyses of variance (ANOVA) with post hoc pairwise comparisons were employed to assess the statistical significance between multiple risk- behaviours and Internet user groups. Moreover, a regression variable plot was conducted to elucidate the linear relationship between the number of hours online per day and the number of risk-behaviours by Internet user group. A critical value of p < 0.05 was considered to be statistically significant for all models.
4.8.5 Statistical procedures for Study V
In study V, the objective was to assess the preventive effect of mental health action in schools (MHAS) on reducing the risk of PIU and psychosocial impairments in adolescents. MHAS was conceptualized by two distinct evidence-based actions used in mental health promotion: (1) education-based action through providing mental health education at school, and (2) gatekeeper-based action through teachers initiating open dialogues with adolescents about their mental health. In order to avoid contamination by students involved in one of the three active interventions in SEYLE, only adolescents participating in the ‘control group’ were included in the analyses. The study sample comprised 2,831 school-based adolescents, in which 47.1 percent were males and 52.9 percent were females (M/F: 1333/1498), with a mean age of 14.83 ± .90 years.
Means, standard deviations and proportions were calculated for sociodemographic variables and stratified by Internet user group. Statistical significance between Internet user groups was measured using Chi-squared (χ²) tests for categorical variables and t-tests for dimensional variables. Baseline prevalence of PIU and psychosocial impairments in adolescents were calculated for the total sample. To explore variations between subgroups, prevalence rates for psychosocial impairments were computed and stratified by gender and Internet user group. Multiple pairwise comparisons assessing the significance of psychosocial impairments by gender and Internet user group were conducted using the two-sided z-test with Bonferroni adjusted p-values. In order to study the preventive effect of mental health actions, adolescents testing positive for PIU and psychosocial impairments at baseline evaluation were excluded from the analysis. By analysing incident cases, this method provided the opportunity to calculate the cumulative incidence and compare absolute risk-reductions (ARR) between students exposed and unexposed to mental health actions implemented in European schools. Statistical significance between adolescents’ exposure levels (exposed vs. unexposed) were evaluated using Fisher’s exact test.
To study the independent effect of education-based action, adolescents who received gatekeeper-based action at school were excluded from the analysis, while the analysis on gatekeeper-based action excluded students previously involved in education-based action at school. In a longitudinal analysis, generalized linear models (GLM) using maximum likelihood estimates (MLEs) were employed to assess the independent effects of education-based action (measured at T0) and gatekeeper-based action (measured at T0) on reducing the onset risk of PIU (measured at T1). Regression coefficients (β) with standard errors (SE) and p-values were reported for GLM models.
The effect of MHAS [i.e. the combined effect of education- and gatekeeper-based actions] on PIU and psychosocial impairments were also examined. In this longitudinal analysis, multilevel mixed-effects logistic regression (MELR) models were conducted to assess the preventive effect of MHAS (measured at T0) on reducing the onset risk of PIU and related psychosocial impairments (measured at T1) in school-based adolescents. Robust variance estimates were applied to the MELR models in order to adjust for the clustered data. Odds ratios with 95% confidence intervals (CI) and p-values are reported for respective models. Both GLM and MELR models were adjusted for baseline sociodemographic variables. A critical value of p < 0.05 was considered to be statistically significant for all models.